README.md

metadata

license: cc-by-4.0

SuperMap: CMU Gates and Scafie annotations

Point clouds and pointwise semantic labels for the SuperMap benchmark. Bounding-box ground truth and evaluation tools are not included.

Files

Archive	Sequences
gates_0511.zip	Static first loop, change loop 1, change loop 2
scafie_0815.zip	Indoor/outdoor

Extract the ZIPs to access each sequence's files:

• point_cloud.txt: one X Y Z R G B point per line.
• point_cloud.zlib: compressed labels, one unsigned byte per point in the same row order.
• label_map.json: class names and colors for that sequence.

manifest.json provides archive checksums and file lists.

Load labels

import zlib
from pathlib import Path

labels = zlib.decompress(Path("point_cloud.zlib").read_bytes())
# labels[i] corresponds to row i of point_cloud.txt (starting at 0).

The Gates archive also includes scripts/visualize_worker_response.py (requires NumPy). It writes a point-cloud PLY with label 0 in red and other points in white.

Notes

• Preserve point order and use each sequence's own class map.
• Class mappings, the meaning of label 255, coordinate units/frames, and the six changed object identities have not been verified for this release. No train/validation/test split is specified.
• In Scafie, one ladder/staircase and one fire extinguisher were labeled as TRASHCAN.

Citation

Please cite the SuperMap paper:

@misc{zhao2026supermap,
  title={SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation},
  author={Shibo Zhao and Guofei Chen and Honghao Zhu and Zhiheng Li and Changwei Yao and Nader Zantout and Seungchan Kim and Wenshan Wang and Ji Zhang and Sebastian Scherer},
  year={2026},
  eprint={2608.22896},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  doi={10.48550/arXiv.2608.22896},
  url={https://arxiv.org/abs/2608.22896}
}

License

CC BY 4.0: reuse and modification, including commercial use, are permitted with attribution, a license link, and an indication of changes. See LICENSE.